Advanced Analytics at the Bank of England...• More data, increased computing power, technical advances –Change of circumstances • Lessons from the financial crisis –Change

Advanced Analytics at the Bank of

England

Presentation to the Riksbank conference on Big Data: Building Strategies

for Central Banks in Light of the Data Revolution

9 September 2015

Outline

• Why are we interested?

• What are we interested in?

– Matched micro data sets

– Text mining

– Visualisation

– Machine learning

• But… there are no free lunches, so what’s the bill?

Advanced Analytics at the Bank of England

2

Why are we interested in Big Data?

• What do we mean by the term?

– Fuzzy meaning, covering data, techniques and attitude


– Change of responsibilities

• The arrival of the PRA

– Change of opportunity

• More data, increased computing power, technical advances

– Change of circumstances

• Lessons from the financial crisis

– Change of philosophy

• Inductive vs deductive reasoning


3


• What do we mean by the term

– Very loose meaning, covering data, techniques and attitude











4














5














6














7

Is Correlation the New Causality?

Karl Popper (Source: http://en.wikipedia.org/wiki/Karl_Popper)

Hal Varian (Source: http://en.wikipedia.org/wiki/Hal_Varian)

15


What are we interested in?

• Gaining a richer understanding of the phenomenon of interest

– Can help disentangle cause and effect…

– …and identify the underlying issue that needs to be addressed

• Getting a speedier reading of developments in the economy and

financial system

– ‘Nowcasting’ and ‘nearcasting’

– This might be particularly important when the system is undergoing

rapid changes

• Quantifying previously purely qualitative data

– Eg text


9

Matched micro data sets


10

Loan-to-income multiple ≥ 4.5

Source: Data are based on the Bank of England’s internal Product Sales Database collected by the FCA.


Sources: WhenFresh (Zoopla listings), Land Registry Price Paid, Land Registry Cash/Mortgage data, FCA Product Sales Data on mortgages, ONS

Postcode Directory.


Sources: WhenFresh (Zoopla listings), Land Registry Price Paid, Land Registry Cash/Mortgage data, FCA Product Sales Data on mortgages, ONS

Postcode Directory.



•14

Highly granular data sets


15

Advanced Analytics at the Bank of England •16

EMIR Data

Positions in

outstanding CHF-

denominated FX

derivatives

positions on

15/1/15

Text Analytics


17

4

5

6

7

8

9

10

Google index of JSA

Unemployment rate

Percent of

labour force

Source: ONS; Google. Notes: The Google indices are mean and variance adjusted to put on the same scale as the unemployment rate and wage growth. The Google

indices are drawn from searches containing the terms “salaries” and “job seekers allowance”. See Mclaren and Shanbhogue (2011) for further details.

Googling the Labour Market

-8

-6

-4

-2

0

2

4

6

8

10

Google index of salaries

Wage growth

Percent year

on year

20


http://www.bankofengland.co.uk/publications/Documents/quarterlybulletin/qb110206.pdf


19


20

Advanced Analytics at the Bank of England 21


22


23

Visualisation


24

Asset class

correlation

heatmap


•26

BoE communication

21


Machine learning


28

Issues with analysing ‘Big Data’

• Example: CPI micro-data

• The ONS has produced a data set comprising:

– 215 months (Feb 1996-Dec 2013)

– ~110,000 prices collected per month (not the same number each

month)

– 1,113 items (not the same items each year)

– 71 COICOP classes

– various other meta-data (eg type of shop, region etc)

– in total: 24,442,988 records with 25 fields

– 611,074,700 pieces of data


29

Issue 1: the stability of annual inflation

21

0

1

2

3

4

5

6

1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2009 2010 2011 2012 2013 2014

UK CPI inflation 12m ratePercentage change over 12 months


Issue 2: over-fitting: what to do?

• T is normally >> N; here N >>T (indeed this is occasionally used

as a definition of ‘big data’ by statistically-minded analysts)

• Aggregate the data? – eg by type of good

– But that tends to obviate the point of using the micro data

• Shrink the dimensionality of the matrix of explanatory variables:

– PCA/factor models extract combinations of the variables that explain

most of the variance of the dependent variable

– So you can keep much of the information that is contained in the

data set without getting into large statistical problems

– A key issue here though is how to interpret the resulting model, and

the components/factors may be unstable over time

• Penalised regressions


31

Issue 3: explaining non-linear functions

21

PLUMBERDAYTIME_HOURLY_RATE

≤106.9386 >106.9386

ORANGECLASS_1EACH

≤104.6805 > 104.6805

CANNED_FISHTUNA180200G

≤102.279 > 102.279

WOMENS_NIGHTDRESSPYJAMAS

≤93.3806 > 93.3806

WINDOWCLEAN_3BED_SEMI

≤101.5122 > 101.5122

WASHING_POWDER_AUTOMATIC

≤99.0384 > 99.0384

DOOR_HANDLEPACK

≤109.8218 > 109.8218


• Try explaining the intuition behind this relationship to busy policy

makers…

Issue 4: Stability

21

0%

2%

4%

6%

8%

10%

12%

14%

16%

18%

20%

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29

% o

f th

e t

ota

l nu

mb

er

of

test

cas

es

Run number

% false positives over 30 random samples

0%

10%

20%

30%

40%

50%

60%

70%

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29

% o

f tr

ue

po

siti

ves

Run number

Positives correctly identified over 30 random samples

• An issue that is closely linked to over-fitting is the stability of the

models

• This is a particularly important issues when there is no strong a

priori reason to think that the world works in this way

• (Though a priori thinking can also be misleading at times)


Issue 5: Confidentiality / ‘Big Brother’ state


• This was not relevant to the CPI work

• In general, the more detailed and granular the data set is, the

more likely it is to contain confidential information

• We must ensure that:

– we only use data for appropriate reasons

– the minimum number of people are able to see any confidential data

given the needs of the situation

– data are stored securely and professionally

35

Issue 6: Practical issues


• Hiring!

• IT!

36

Conclusion

• Do these issues mean that ‘Big Data’ is likely to be a passing

fad?

– No!

• The data exist and the BoE has the responsibility and opportunity

to use them to help us understand economic developments and

the structure of the economy and financial system

• But the issues do mean that this is no panacea

• Just as with any other empirical work the data need to be

cleaned and understood (both of which are more difficult with

larger data sets) prior to analysis

• And then analysed carefully using appropriate methods


37